The Competitive Edge of deAI
Decentralised AI isn't just an ethical paradigm, its a business advantage.
May 2026 · 2,200 words ·
The standard argument for decentralised AI goes something like this: centralised platforms are bad because they're controlled by corporations, so we should build open alternatives because freedom is good. That argument is fine. It is also losing.
You don't win a technology transition by being morally correct. You win by being structurally better. And that's what I think most people are missing about where decentralised AI actually sits right now.
I've spent the last few years building Lilypad, a distributed compute network for AI inference and ML pipelines. We closed the company in September 2025 and open-sourced everything. Before that happens to the argument I'm about to make, let me write it down properly.
Decentralised AI infrastructure doesn't just offer an alternative to centralised platforms. It solves specific structural problems that centralised platforms cannot solve without undermining their own business model. That's a different claim, and a stronger one.
Here are the seven advantages I believe are durable.
1. Native Payment Rails
Centralised AI platforms are built for credit cards and enterprise invoicing. That design choice seems trivial until you try to get a GPU job run by a researcher in Kenya or compensate a model contributor in Vietnam.
The actual problem isn't payments technology. It's that centralised platforms have no incentive to fix cross-border payment friction because their customers are mostly enterprises in a handful of high-income countries. The addressable market they've optimised for is already captured.
Decentralised infrastructure has crypto settlement baked into the protocol layer. Not bolted on, not a feature request. The payment rail and the compute rail are the same rail.
CENTRALISED PLATFORM DECENTRALISED PROTOCOL
───────────────────── ──────────────────────────────
User User
│ │
▼ ▼
Credit card / enterprise invoice Web3 wallet OR Stripe fiat
│ │
▼ ▼
Platform (takes cut, sets terms) Smart contract (automated,
│ transparent, no intermediary)
▼ │
GPU provider (US/EU only) ▼
GPU provider (anywhere on earth)
The practical implication: a decentralised network can source compute from nodes in markets where centralised platforms have no reach, at prices those markets make viable. That is not a philosophical point. It is a procurement advantage.
2. Provenance Pipelines
When an AI model produces an output, three questions matter enormously and are currently unanswerable on most platforms:
- What training data produced this behaviour?
- Who contributed the model weights that generated this result?
- If this model was fine-tuned on another model, what's the full lineage?
These are not niche concerns for AI ethicists. They are the questions that will determine legal liability as AI regulation matures, royalty flows as model economies develop, and scientific reproducibility as AI research compounds.
Centralised platforms hold all of this information and share essentially none of it. Not because they're malicious (mostly). Because their business model is predicated on opacity: you pay for the output, not the understanding.
Cryptographic job verification and on-chain records make provenance traceable by design. Every job execution leaves a verifiable record. Model lineage is auditable. When a fine-tuned model earns revenue, the attribution chain is visible.
MODEL PROVENANCE CHAIN (decentralised)
Foundation model (open weights)
│
├── recorded on-chain: creator, licence, version hash
│
Fine-tune (domain-specific)
│
├── recorded on-chain: who fine-tuned, what data, what model version
│
Deployment
│
├── job execution: cryptographically verified
│
Revenue
│
└── attribution flows back up the chain automatically
This is the infrastructure layer for a functioning model economy. Without it, you're trusting a platform's internal accounting. That trust will be tested.
3. Permissionless Global Participation
This one sounds ideological. Bear with me, because it has a hard economic implication.
Centralised platforms have onboarding processes, know-your-customer requirements, geographic restrictions, and terms of service that exclude large categories of potential contributors. A GPU node operator in Southeast Asia with a competitive rig can't necessarily sell compute to AWS. A researcher at a small African university may not have institutional credentials for enterprise AI access. A solo developer building an open-source AI tool may not qualify for the pricing tier where the interesting capabilities live.
Permissionless participation means those people are in the network. Not as an act of charity. As an act of market expansion.
The diversity argument for decentralised AI is usually framed around values (more voices, more perspectives, better outcomes). That argument is real. But there's a prior argument: the centralised platform is systematically under-indexing on available talent and compute. Permissionless infrastructure captures what the credentialled system rejects.
PARTICIPATION MAP
Centralised platform
────────────────────
Geography: US / EU / select markets
│
Participation type: Enterprise contracts
Individual accounts (KYC)
│
Barriers: Payment method
Geographic restriction
TOS approval
Pricing tier
Decentralised protocol
──────────────────────
Geography: Any node with internet access
│
Participation type: Wallet + job = participation
│
Barriers: Hardware (compute)
Network access
Full-stack open interfaces (CLI, API, GUI) make this concrete. The protocol doesn't care where you are or who you are. It cares whether you can run the job or provide the compute.
4. Fair Creator Economics
The model economy has a credit problem. When a model creator's weights are incorporated into a commercial product, the creator typically receives: nothing. When a fine-tune built on open weights goes on to generate millions in revenue, the original contributors see: nothing.
This isn't a new complaint. But the mechanism for fixing it has only recently become practical.
On-chain revenue distribution means model owners set pricing, earn per use, and receive rewards directly. The smart contract executes the distribution. There is no platform in between deciding what's fair.
The comparison to the music industry is instructive and worth being precise about. Streaming platforms created a mechanism for distributing royalties but retained control over the rate-setting. The result was a system that was technically fairer than piracy but structurally favoured platforms over creators. Decentralised creator economics removes the rate-setter.
REVENUE FLOW COMPARISON
Centralised model marketplace
─────────────────────────────
User pays → Platform → Platform decides split → Creator (residual)
→ Platform retains significant margin
→ Terms can change unilaterally
Decentralised model marketplace
────────────────────────────────
User pays → Smart contract → Creator (set rate, automatic)
→ Compute provider (set rate, automatic)
→ Protocol treasury (transparent, fixed)
The claim here isn't that decentralised economics are perfect. It's that they're auditable. You can see what you agreed to and verify it was honoured.
5. Composable Infrastructure
The AI stack is not a single product. It is a pipeline: data ingestion, storage, processing, model training, inference, output routing, agent orchestration. Every serious AI deployment involves multiple systems talking to each other.
Centralised platforms want to own as much of this pipeline as possible. That is rational from their perspective. From the builder's perspective, it creates lock-in at every layer.
Composable infrastructure is designed for interoperability by default. Each component exposes standard interfaces. You can swap the storage layer (Filecoin, IPFS, Arweave) without rewriting the inference layer. You can connect agent frameworks (Olas, ElizaOS) without negotiating enterprise contracts.
COMPOSABLE AI STACK
Agent layer Olas / ElizaOS / custom agent
│
Orchestration Pipeline runner (open interfaces)
│
Inference Distributed compute network
┌────┴────┐
Storage Filecoin Arweave IPFS
│
Data Vana / open datasets / contributor pools
The network effects here run differently from centralised platforms. Centralised platforms grow by adding users inside a walled garden. Composable infrastructure grows by adding integrations that each bring their own user bases. The value accrues to the protocol, not to a single company's platform.
6. Censorship Resistance
I want to be precise about this one, because it's often argued badly.
The claim is not that all content restrictions are wrong. It is that the entity making the restriction decision should not also be the entity that profits from the infrastructure. When a centralised platform deplatforms a model or restricts a use case, it is simultaneously acting as publisher, regulator, and commercial operator. Those roles have a conflict built in.
The specific cases where this matters:
Open scientific research. Models trained on genomic data, epidemiological records, or chemical compounds can trigger content policies designed for consumer contexts. Research that should be freely publishable gets blocked by safety filters calibrated for a different threat model.
Politically sensitive tools. Fact-checking tools in authoritarian contexts, legal aid tools in countries with hostile governments, investigative journalism applications. All of these are legitimate uses that face real deplatforming risk on commercial infrastructure.
Grassroots innovation. The more mundane case: a developer building something the platform hasn't thought of yet, that doesn't fit the approved use-case categories, that gets flagged by an automated system and has no appeal mechanism.
Fully on-chain execution and settlement means job provenance is verifiable and tamper-proof. No single authority can retroactively alter the record of what ran or why.
This matters less for the median AI application. It matters enormously for the applications where it matters.
7. Designed for the Actual Direction of Travel
The proprietary AI moat argument assumed that model quality was the primary differentiator and that frontier model access required massive centralised compute. Both assumptions are weakening.
Open weights models are closing the capability gap with proprietary frontier models faster than the proprietary labs expected. Fine-tuning on domain-specific data produces specialised models that outperform general-purpose frontier models on their target tasks. Inference is getting cheaper by roughly an order of magnitude every 18 months.
The AI economy is shifting towards:
FROM TO
────────────────────────── ──────────────────────────────────
One model for everything Custom, fine-tuned, specialised models
Human-in-the-loop workflows Agentic, automated pipelines
Cloud-hosted centralised User-owned, locally-run infrastructure
Platform-controlled economics Creator and contributor economics
Opaque provenance Auditable, on-chain lineage
Centralised platforms can participate in parts of this shift. But they cannot embrace the full direction of travel without undermining their own margins. A platform whose revenue model depends on being the compute intermediary cannot also be the platform that most aggressively enables user-owned compute. The incentives don't align.
Decentralised infrastructure was designed for the destination, not the starting point.
The Honest Caveat
I've argued the structural case for decentralised AI as strongly as I can. It's the case I believe. It's also incomplete without the honest counterargument.
Centralised platforms have a massive head start in developer experience, reliability, compliance infrastructure, and enterprise sales. For most AI applications today, the path of least resistance is AWS, Azure, or a frontier model API. That's a real advantage, not just marketing.
The structural advantages I've described are long-term advantages. They compound over time as the AI economy matures, as regulation increases provenance requirements, as model economics develop, as the geographic distribution of AI participation expands. They are not advantages you feel in your first sprint.
The honest prediction: centralised and decentralised infrastructure will coexist for a long time. The question is which direction the equilibrium shifts over a five to ten year horizon, and which layer of the stack captures the most value.
I think the structural arguments above are why decentralised infrastructure wins that longer game. The AI economy that's coming is too global, too diverse, too economically complex, and too legally consequential to be governed by a handful of platform terms-of-service.
What I'd Build Next
If I were starting Lilypad again with what I know now, I'd focus on the provenance pipeline problem first. It's the advantage that becomes more valuable as AI regulation tightens, and it's the problem that centralised platforms are structurally least able to solve. The audit trail for AI outputs is going to matter enormously within three years.
The payment rails and creator economics advantages are real but require network scale to become compelling. Start with provenance. The rest compounds.
Lilypad closed in September 2025. All infrastructure is open-sourced. The argument above is mine, not a pitch.
Related reading: The Decentralised AI Landscape